Chinese semiconductor industry

Status
Not open for further replies.

tokenanalyst

Brigadier
Registered Member

Two semiconductor projects signed in Chun’an, Zhejiang​


A total of 12 projects were signed at the first "Harmonious Symbiosis of Man and Nature" Qiandao Lake Conference recently, with a total investment of approximately 14 billion. These include the Semiconductor Industrial Park Project and the Jingfeng Mingyuan Semiconductor Project.

It is reported that the main investor in the Semiconductor Industrial Park project is Ningbo Lingju Venture Capital Partnership (Limited Partnership), with a total investment of 1 billion yuan and a planned land area of about 58 acres. It will invest in the construction of a Semiconductor Industrial Park, including semiconductor and storage R&D and production bases, experts ( Academician) workstation and product R&D laboratory and display center. Within five years after it is completed and put into operation, the annual gross industrial output value will be 1 billion yuan, and the annual research and experimental development (R&D) investment will be 20 million yuan.

The Jingfeng Mingyuan Semiconductor project has a total investment of 100 million yuan. It is a subsidiary of Shanghai Jingfeng Mingyuan Semiconductor Co., Ltd. and is positioned in the R&D and sales of high-end power management chips for "polyphase digital high-power power management chips and intelligent power devices". The successful launch of the product is of great value to the smooth promotion of domestic CPUs. It intends to lease 2,000 square meters of office space and 400 square meters of laboratory space in Qiandao Lake Zhigu Building, with about 70 employees. The Hangzhou company expects operating income of 15 million yuan in 2024.​
 

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
Some updates from HQ weibo account
华为麒麟5G平台快落地中端系列了,只是量不大,供小于需。另外中端系列也安排了竖向小折叠,小折双品牌线+外折叠+内折叠。直板旗舰系列影像硬件将迎来大升级
一个行业好消息,1.5K屏下前摄方案已开案,明年见
CMOS被国产逐步攻克以后,下一步就是柔性玻璃,我说的

Alright, so 5G Kirin chip for medium phones will come out soon. Will be used in foldable phones also. But the most obvious sign will be Nova 12.

Flagship pad CIS hardware to get major upgrade

CMOS is the current major domestication part.

Next one is flexible OLED

btw, personally speaking. Phones and EVs represent two of the hardware requiring the most demanding chips. If they can show reasonable progress in moving to domestic suppliers for these areas, then they are a good to pushing foreign suppliers out of other products too
 

gelgoog

Lieutenant General
Registered Member
Please, Log in or Register to view URLs content!

November 8, 2023

If the U.S. government intends to curb China’s adoption of emerging RISC-V architecture to develop homegrown chips, it may be getting late.
Last month, China’s Shandong University deployed a server cluster with RISC-V CPUs. The system has a total of 3,072 cores, with 48 nodes of 64-bit RISC-V CPUs.
This is the first commercial implementation of a RISC-V cluster in the cloud, claimed David Chen, director of ecosystem at Alibaba, during a presentation at the RISC-V Summit in Santa Clara, California.
The system is for Shandong University’s teaching and research purposes but is also available for cloud computing instances, Chen said in response to a question from HPCwire for more details on the system. He added that the system was finished and delivered in September and October.
...
The RISC-V system delivered to Shandong University system uses Sophgo’s SG2042 chip, which has a clock speed of 2GHz and 64 M.B. of cache. The system supports the PCIe Gen 4 interface.
...

Please, Log in or Register to view URLs content!
...
Leveraging the RA-JAPerf benchmarking suite, we discover that on average, the SG2042 delivers, per core, between five and ten times the performance compared to the nearest widely available RISC-V hardware. We found that, on average, the x86 high performance CPUs under test outperform the SG2042 by between four and eight times for multi-threaded workloads, although some individual kernels do perform faster on the SG2042.
...
The Sophon SG2042 CPU is a 64-core processor running at 2GHz and organised in clusters of four XuanTie C920 cores. Each 64-bit core, designed by T-Head, is designed for high performance and adopts a 12-stage out-of-order multiple issue superscalar pipeline design. Providing the RV64GCV instruction set, the C920 has three decode, four rename/dispatch, eight issue/execute and two load/store execution units. Version 0.7.1 of the vectorisation standard extension (RVV v0.7.1) is supported, with a vector width of 128 bits supporting data types FP16, FP32, FP64, INT8, INT16, INT32, and INT64. Each C920 core contains 64KB of L1 instruction (I) and data (D) cache, 1MB of L2 cache which is shared between the cluster of four cores, and 64MB of L3 system cache which is shared by all cores in the package. The SG2042 also provides four DDR4-3200 memory controllers, and 32 lanes of PCI-E Gen4.
 
Last edited:

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
Please, Log in or Register to view URLs content!

November 8, 2023

If the U.S. government intends to curb China’s adoption of emerging RISC-V architecture to develop homegrown chips, it may be getting late.
Last month, China’s Shandong University deployed a server cluster with RISC-V CPUs. The system has a total of 3,072 cores, with 48 nodes of 64-bit RISC-V CPUs.
This is the first commercial implementation of a RISC-V cluster in the cloud, claimed David Chen, director of ecosystem at Alibaba, during a presentation at the RISC-V Summit in Santa Clara, California.
The system is for Shandong University’s teaching and research purposes but is also available for cloud computing instances, Chen said in response to a question from HPCwire for more details on the system. He added that the system was finished and delivered in September and October.
...
The RISC-V system delivered to Shandong University system uses Sophgo’s SG2042 chip, which has a clock speed of 2GHz and 64 M.B. of cache. The system supports the PCIe Gen 4 interface.
...

Please, Log in or Register to view URLs content!
yeah, it's a pretty nice system.

SG2044 will come out next year and use cores designed by T-Head IIR, which uses the latest RISC-V 1.0 ISA

SG2044_Aug2023.jpeg


should be using C908, which supports Vector 1.0. Vector 1.0 has significantly improved AI performance (you can see the C908 advantage vs C906 in image classification, Visual Wake words & keyword spotting)
Please, Log in or Register to view URLs content!
 

gelgoog

Lieutenant General
Registered Member
SG2044 will come out next year and use cores designed by T-Head IIR, which uses the latest RISC-V 1.0 ISA
...
should be using C908, which supports Vector 1.0. Vector 1.0 has significantly improved AI performance (you can see the C908 advantage vs C906 in image classification, Visual Wake words & keyword spotting)
No, this should be a different core than the C908. That uses dual issue 9-stage in-order pipelines. That is meant for IoT applications.

Since this would be a replacement for the SG2042 it will likely use an out-of-order core. The SG2042 uses the C920. That used a quad issue 12-stage out-of-order pipeline. Looking at that slide the SG2044 likely uses a similar processor core but will have the RISC-V Vector 1.0 extension. And the processor has like twice the memory controllers. It might also have a wider vector unit.

The difference in core design between the C908 and whatever core this uses should be similar to the difference between ARM's Little and Big cores, or Intel's Efficiency and Performance cores.
 
Last edited:

latenlazy

Brigadier
transistor density is probably the only PPAcT metric we can determine without having actual chip to test performance, power consumption, and even cost. It IS the lone Apple to Apple benchmark of anyone’s given process capability.
As we have discussed before, just because you can’t measure transistor performance and power consumption directly that does not mean they are dismissible parameters when in fact they and not transistor density are the most important parameters for a process node. I don’t know how many chip designers would choose say 10% worse switching performance and efficiency for 10% more density but I can’t imagine it’s many. Actual performance is about your realized net outcomes and not an apples to apples abstraction that excludes the most critical determinants of performance. No one is going to accept an apples to apples density comparison when an old 7 nm node has better real world power consumption than another 5 nm node. Why do you think Samsung struggled to get customers?
 
Last edited:

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
If you are like me, you probably ask the question of when more Chinese chips can fulfill domestic demand? And I'm talking about mature node, not advanced nodes, since that's dependent on more things.

Well, I think the good place to look at are the phone & EV supply chain, since these have some of the highest requirements for MCUs, memory, sensors, analog and such.

In Sep of last year, GigaDevice (who is not small by any means) released its first auto grade GD32 series MCUs
Please, Log in or Register to view URLs content!


It was only a year later did they get ASIL-D certification for development process

And I was looking up its delivery process and saw this from April
在存储产品上,公司 NOR Flash、NAND Flash 均推出了全系列的车规 产品,在车载应用领域进展非常顺利,累计出货量已超过 1 亿颗,NAND Flash 车 规产品也在积极进行市场推广。MCU 产品上去年第三季度推出了第一颗车规产 品,主要是针对车身域应用,目前在很多车身域关键应用领域在一些头部客户已 经 DESIGN IN。车规 MCU 产品市场开拓周期比较长,DESIGN IN 后还有可靠性 测试等环节,然后才会正式在客户量产并实现放量,这个时间大概会需要两年左 右。
So basically even a firm with good reputation in China and plenty of experience delivering auto chips (for NAND/NOR Flash) has to slowly work with customers to get its MCU designed in and then go through testing before mass production. And they estimate the entire process to take 2 years

Definitely not easy. So as we run into a situation where auto grade MCU demand has surged, the investment that domestic fabless put into auto chip might still take a while to work out

From July
Please, Log in or Register to view URLs content!
SPI NOR Flash方面,已实现2Gb容量产品量产,目前产品制程以55nm为主,制程上正向40nm推进,产品结构升级趋势明显
司SPI NAND Flash和Parallel NAND Flash两类产品以SLC颗粒为主,目前已实现1Gb-8Gb容量覆盖,产品工艺制程已达到24nm,正进一步向19nm推进。同
Assuming SMIC is doing all the fabbing. They basically still mostly do 55nm for NOR flash, but moving more to 40nm now that SMIC supports it.
For NAND, looks like SMIC is finally getting 19nm Nand process?
From SMIC first half report,
Please, Log in or Register to view URLs content!

one of the things they developed was 4xnm NOR Flash. So it seems to me that GigaDevice moves as SMIC moves

在汽车应用领域,发布GD32A503系列车规级MCU产品,为车身控制、车用照明、智能座舱、辅助驾驶及电机电源等场景提供开发之选。目前,兆易创新MCU产品的工艺制程集中在55nm及40nm
Also saw this, so they are basically concentrating most of their MCU efforts on 55/40nm process

All this tells me if we see good chunk of new Chinese EVs are using domestic MCUs and other chips, then there is a good chance they have eaten into foreign firms market share in other sectors also, since getting into EV supply chain takes so long.

And SMIC progress in mature process is really important for China's domestic chip drive, because firms like GigaDevice needs SMIC to continue to make progress
 

KYli

Brigadier
Please, Log in or Register to view URLs content!

President Joe Biden has adopted a two-pronged approach to constrain China’s high-tech progress, curbing Beijing’s access to leading-edge chips while bolstering semiconductor production in the US.

He’s about to ratchet up the pressure further, shifting focus to an emerging arena of the contest for technological supremacy: the process of packaging semiconductors that’s increasingly seen as a path to achieving higher performance.

Only the US isn’t alone is recognizing the potential of so-called advanced packaging: China, too, is capitalizing on an area that isn’t subject to sanctions, capturing global market share and achieving progress denied it in manufacturing high-end chips.

“Packaging is the new pillar of innovation in the semiconductor industry – it will change the industry drastically,” said Jim McGregor, founder of technology analysts Tirias Research. For China, which doesn’t yet have state-of-the-art capabilities, “it’s definitely easier for them to ramp up” here, since it isn’t restricted by the US government. “Packaging could help them bridge the gap,” he said.

Up until very recently, the business of packaging semiconductors – encasing chips in materials that both protect them and connect them to the electronic device they’re part of – was, at best, an afterthought for the industry. So it was outsourced, mainly to Asia, with China a prime beneficiary: today, the US accounts for just 3% of the world’s packaging capacity, according to Intel Corp.


Yet suddenly, advanced packaging is everywhere: Intel is banking on it as a core part of the US chip giant’s strategy to return to competitiveness; China sees it as a means of building out domestic semiconductor capacity; and now Washington is turning to it as part of its own plans for self-sufficiency.

More than a year after the CHIPS and Science Act came into being, the Biden administration has outlined plans for a $3 billion National Advanced Packaging Manufacturing Program, after recently tapping a director for the center. The goal is to create multiple high-volume packaging facilities by the end of the decade, said Under Secretary of State of Commerce Laurie Locascio — and reduce reliance on Asian supply lines that pose a security risk the US “just can’t accept.”

Read more: US Launches $3 Billion Effort to Boost Advanced Chip Packaging

With advanced packaging rapidly becoming a new front in the global conflict over chips, some argue it’s long overdue.

The administration has until now focused on subsidies to bring back chipmaking to the US, but “we can’t ignore packaging because you can’t do one without the other,” said Representative Jay Obernolte, a California Republican who is one of two vice-chairs of the Congressional Artificial Intelligence Caucus. “It wouldn’t matter if we did 100% of our chip manufacturing onshore if the packaging is still offshore,” he added. The White House didn’t respond in time for publication when asked for comment on advanced packaging.

Assembly, testing and packaging – usually considered together as “back-end” manufacturing - was always the least glamorous end of the semiconductor industry, with less innovation and lower added value than the “front end” business of making chips with features measured in the billionths of a meter. Yet the level of sophistication is rising fast as new technologies enable chips to be combined, stacked and their performance enhanced in what industry executives are calling an inflection point.

Advanced packaging can’t help China compete with leading-edge semiconductor developments from the U.S., but it allows Beijing to build faster, cheaper systems for computing by stitching together different chips closely together. In that case China could save its latest chip technology, which is expensive and likely available in limited volume, for the most important part of the chip and use older, cheaper technologies to make chips that carry out other functions like battery management and sensor controls, combining the whole in a powerful package.

It's a “pivotal solution,” said Bloomberg Intelligence technology analyst Charles Shum. “It doesn’t merely enhance chip-processing speed but crucially enables seamless integration of varied chip types.” As a result, he said, it’s “set to reshape the semiconductor-manufacturing landscape.”

Beijing has long made a strategic priority of semiconductor packaging technologies, including in President Xi Jinping’s Made in China program announced in 2015. China has 38% of the world’s assembly, testing and packaging market, the most of any nation, according to the US-based Semiconductor Industry Association. While it lags behind Taiwan and the US in advanced technology, analysts agree that unlike in wafer processing, it’s in a much better position to be able to catch up.

China already boasts the most back-end facilities by number, including the world’s third-largest assembly and testing company, JCET Group, which trails only Taiwan’s ASE Group and Amkor Technology of the US in revenue. What’s more, Chinese companies are building market share, including through JCET’s acquisition of an advanced facility in Singapore and construction of an advanced packaging plant in its hometown of Jiangyin.

“For China, one way around technology transfer restrictions is advanced packaging, because so far it’s a safe space that everyone invests in,” said Mathieu Duchatel of the Institut Montaigne think tank, a Taiwan-based China expert who studies the geopolitics of technology.


It’s a realization now touching Washington as it seeks to deny Beijing access to the kind of advanced computing technologies that could be put to military use – with questionable success.

When Huawei Technologies Inc. quietly released its Mate 60 Pro smartphone in September, China hawks in Washington raised questions as to why US export controls had failed to prevent a development supposedly beyond Beijing’s capabilities.

In testimony to the House Sept. 19, Commerce Secretary Gina Raimondo defended the Biden administration’s focus on denying China access to leading-edge chips and the equipment to make them. But she was primed on advanced packaging. The US needs to ramp up its own advanced packaging capacities, she said, since “chips can only get so small, which means all the special sauce is in the packaging.”

One reason for the sudden focus on that special sauce is its necessity to the kind of high-power semiconductors needed for artificial intelligence applications. Indeed, a shortage of a particular type of packaging known as Chip on Wafer on Substrate, or CoWoS, is a key bottleneck in the production of Nvidia Corp’s AI chips.

Taiwan Semiconductor Manufacturing Co., the main chipmaker for Nvidia, this summer committed $3 billion to a packaging plant to help alleviate the blockage. CEO C.C. Wei told investors on the company’s third-quarter earnings call that the company planned to double CoWoS capacity by the end of next year.

While TSMC has been working on the technology for 12 years, it was a niche application that only took off this year, Jun He, Vice President of Advanced Packaging Technology, told a conference in Taipei in October. “We’re building capacity like crazy,” said Jun He, adding that “everybody, probably even in Starbucks,” is talking about CoWoS.

It’s not just TSMC. Micron Inc. is setting up a $2.75 billion back-end facility in India, while Intel agreed to build a $4.6 billion assembly and test plant in Poland and is putting some $7 billion into advanced packaging in Malaysia. South Korea’s SK Hynix said last year that it plans to invest $15 billion in a packaging facility in the US.

Intel has “some very unique technology now in the packaging area,” Chief Executive Officer Pat Gelsinger said in an interview. “Everybody who’s doing AI chip work today is looking to say, wow, this is the way that I can advance my AI chip capabilities.”

That has some analysts predicting a bonanza for companies in the sphere. According to McKinsey, high-performance chips for data centers, AI accelerators, and consumer electronics will create the greatest demand for advanced packaging technologies.

The number of chips shipped that use advanced packaging is forecast to increase tenfold in the next 18 months – but that could soar to 100 times if it becomes standard in smartphones, Jeffries analysts Mark Lipacis and Vedvati Shrotre wrote in a Sept. 14 report that classed the technology as part of a “tectonic shift” in the industry.

The reason, alluded to by Raimondo, is that chipmaking is running up against the limits of physics.

Chips have been getting better over the last fifty years in large part through advances in production technology. The components now contain up to tens of billions of the tiny transistors that give them the ability to store or process information. But now that path of advancement, called Moore’s Law after Intel’s founder, is coming up against fundamental barriers that are making improvements more difficult and vastly expensive to achieve.

Moore’s Law – more of an observation – states that the number of transistors on a chip doubles about every two years. As that pace of progress slows, and companies “are not able to deliver twice the transistors, at half the cost, at twice the clock speed, and at lower power levels every two years, the industry has begun to rely more on advanced packaging techniques to pick up the slack,” Lipacis and Shrotre wrote.

Instead of cramming ever more tiny components on to one piece of silicon, many designers and companies are touting the benefits of a modular approach, of building products out of several “chiplets” tightly packed together in the same package.
 
Status
Not open for further replies.
Top